Method for screening and optimizing audio-frequency keyword template

Abstract

The invention provides a method for screening and optimizing an audio-frequency keyword template. The method comprises the steps of step1, extracting features of each audio-frequency keyword template sample, making extracted features through a deep neural network and calculating posterior probabilities of all phonemes on a given phoneme set; step2, calculating posterior probability stability scores, pronunciation reliability scores and neighborhood similarity scores of the template; step3, calculating a weighted average of the above three scores of each audio-frequency keyword template and denoting the weighted average as an average score; step4, sequencing according to the order of average scores from large to small and selecting the first L audio-frequency keyword templates as representative pronunciation templates; step5, processing each representative pronunciation template, adjusting the posterior probability of each pronunciation unit on each frame of the pronunciation sequence and minimizing the neighborhood similarity scores of the template; generating optimized L audio-frequency retrieval word templates.

Description

technical field [0001] The invention belongs to the field of speech recognition, and in particular relates to a method for screening and optimizing audio keyword templates. Background technique [0002] The keyword retrieval task is to quickly find the location of a given keyword from large-scale and diverse speech data. In the keyword retrieval task based on speech clips, keywords to be retrieved are given in the form of a set of audio clip templates. These fragments are usually from different speakers or extracted from different contexts, and thus differ in the information they contain. In order to obtain retrieval results with better generalization, that is, in order to be able to process keywords from different speakers or with different contexts in the speech to be retrieved, it is necessary to make full use of as many audio clips as possible of a certain keyword. The traditional approach is to average all the templates belonging to a single keyword to obtain a single...

AI Technical Summary

Problems solved by technology

[0003] However, in practical tasks, different audio clips of keywords often have large differences in quality, which may come from factors such as noise, channel mismatch, labeling errors, etc.

Such audio clips may not be sufficiently discriminative, and thus may lead to poor retrieval performance of the system if directly introduced into the keyword retrieval process

Images